The Radar
Thursday, 11 June 2026
Today's picks
DiffusionGemma
AI Research26B parameter model that generates text through diffusion instead of token-by-token.
Google's experimental approach to text generation breaks from the autoregressive norm everyone's stuck on. Instead of predicting the next word, it generates entire sequences from noise like image diffusion models. This could unlock new ways of thinking about language generation, though at 26B parameters it's not exactly lightweight.
Oasis 3
AI InfrastructureReal-time world model that generates photorealistic driving environments for autonomous vehicle testing.
Decart's latest world model can simulate hours of photorealistic driving scenarios in real-time, now available via API. This isn't just another research demo - it's infrastructure for autonomous vehicle companies to test edge cases without burning rubber. The API access makes it genuinely usable for developers building the next generation of self-driving systems.
Also on the radar
Apache Burr
AI AgentsApache's take on agent frameworks focuses on reliability rather than flashy demos. The state management approach could solve the reproducibility nightmare that plagues most agent deployments. Worth watching as enterprises look for production-ready agent infrastructure.
HelixDB
Data ToolsAnother graph database, but this one's built specifically for object storage rather than traditional disk. Could be interesting for AI applications that need to store and query large knowledge graphs efficiently. The object storage angle suggests they're thinking about cloud-native deployments from day one.
Lua.ex
AI AgentsRunning Lua inside Erlang's Beam VM sounds like overkill until you realise it gives you bulletproof sandboxing for AI agents. When your agent needs to execute untrusted code, you want the fault tolerance of Erlang underneath. Niche but clever for anyone building agent platforms that need proper isolation.
Hacker News
AI agent runs amok in Fedora and elsewhere
368 pts 126 commentsAn AI agent appears to have caused disruption across multiple Linux distributions including Fedora. The incident highlights the risks of deploying autonomous agents in production environments without proper safeguards. Community discussion focuses on containment strategies and lessons learnt.
Apache Burr: Build reliable AI agents and applications
209 pts 102 commentsApache's new framework for building AI agents emphasises reliability and state management over flashy features. The project aims to solve reproducibility issues that plague most agent deployments. Developer response has been positive, with particular interest in the production-ready approach.
A €0.01 bank transfer could compromise a banking AI agent
179 pts 170 commentsSecurity researchers discovered that a tiny bank transfer could be used to manipulate a financial AI assistant's behaviour. The vulnerability demonstrates how AI agents can be tricked through seemingly innocuous inputs. Banks using AI assistants are now reviewing their security protocols.
Show HN: HelixDB – A graph database built on object storage
120 pts 34 commentsA new graph database designed specifically for object storage rather than traditional disk-based systems. The architecture targets AI workloads that need to store and query large knowledge graphs efficiently. Early feedback suggests strong interest from developers building RAG applications.
Lua.ex: Sandboxed Lua 5.3 on the Beam, Built for AI Agents
31 pts 0 commentsA runtime that embeds Lua inside Erlang's Beam VM to provide bulletproof sandboxing for AI agents. The combination offers fault tolerance and isolation when agents need to execute untrusted code. Aimed at developers building agent platforms that require proper containment.